Abstract Generated abstract
The paper studies an assembly model in queueing theory with several independent Poisson input streams, where completed units require one element from each stream and the remaining elements form a queue vector. Using barycentric coordinates on a simplex lattice, it defines functions that express the probabilities of queue states and that reduce, in the two-stream case, to cylindrical Bessel functions. The author derives a generating function, addition and differentiation formulas, Fourier and integral representations, and asymptotic estimates for these generalized functions in regimes relevant to large arrival intensities.
Full Text
UDC 517.516
MATHEMATICS
Corresponding Member of the Academy of Sciences of the USSR L. A. LYUSTERNIK
ON A PROBLEM IN QUEUEING THEORY AND A RELATED GENERALIZATION OF THE CYLINDRICAL FUNCTIONS \(J_k\)
1. It will be convenient for us to use coordinates \(y_j,\ j=0,1,\ldots,n\), in the \(n\)-dimensional Euclidean space \(E_n\), which we shall call barycentric (abbreviated b.c.) (in what follows \(\sum_j\) and \(\prod_j\) denote \(\sum_{j=0}^n\), \(\prod_{j=0}^n\)). Let the unit vectors \(e_j,\ j=0,1,\ldots,n\), be the vertices of a regular \(n\)-dimensional simplex \(D_n\), whose center lies at the origin \(O\). Every vector \(Y \subset E_n\) will be represented in the form \(Y=\sum_j y_j e_j=[y_j]\), where \(y_j\) are b.c. of the vector \(Y\). Since, for any \(y\), \([y_j]=[y_j+y]\), one may fix one of the b.c. or their sum. For \(\sum_j y_j=1\) we obtain the usual b.c. Setting
\[
y_{\mathrm{cp}}=\frac{1}{n+1}\sum_j y_j,
\]
we obtain the canonical b.c. \(y'_j\) of the vector \([y_i]=[y'_j]\),
\[
y'_j=y_j-y_{\mathrm{cp}},\qquad \sum_j y'_j=0.
\]
Denoting \(\bar y=\min_j y_j,\ \bar y_j=y_j-\bar y;\ \bar y_j\ge 0,\ \min \bar y_j=0\), we shall call \(\bar y_j\) the second canonical b.c. of the vector \([y_j]=[\bar y_j]\). For \(Y=[y_i]=[y'_j]\) we have
\[
\|Y\|^2=\frac{n+1}{n}\sum_j y_j^2-\frac{1}{n}\left(\sum_j y_j\right)^2=\frac{n+1}{n}\sum_j y_j'^2=
\]
\[
=\frac{n+1}{n}\sum_{k=1}^n \frac{k}{k+1}\left[y_k-\frac{1}{k}\sum_{i=0}^{k-1}y_i\right]^2.
\tag{1}
\]
By \(\Omega_n\) we shall denote the lattice of integer vectors \(K=[k_j]\) in \(E_n\), where \(k_j\) are integers; \(W_n\) is the Dirichlet region of the lattice \(\Omega_n\), its volume
\[
|W_n|=\sqrt{(n+1)^{\,n-1}/n^n}.
\]
For a function \(f(Y)=f([y_j])\) defined in \(E_n\),
\[
\sum_j \frac{\partial f}{\partial y_j}=0;\qquad
\sum_j \frac{\partial^2 f}{\partial y_j^2}=\frac{n+1}{n}\Delta f=\frac{n+1}{n}\sum_{i=1}^n \frac{\partial^2 f}{\partial x_i^2}
\tag{2}
\]
(\(x_i\) are Cartesian coordinates).
2. The assembly problem. A generalization of the problem of taxis and passengers \((^1)\) is the assembly problem. Given are \((n+1)\) streams \(P_j,\ j=0,1,\ldots,n\). The probability that by time \(t\) there arrive \(k\) elements of the stream \(P_j\) is equal to \(p_j(k,t)\). We shall regard the streams as Poisson:
\[
p_j(k,t)=e^{-\lambda_j t}\frac{(\lambda_j t)^k}{k!}\qquad \left(\frac{1}{k!}=0\ \text{for } k<0\right).
\tag{3}
\]
Abstracting from the assembly time, we assume that if by time \(t\) there have appeared \(k_j\) elements of the stream \(P_j,\ j=0,1,\ldots,n\), then \(k=\min_j k_j\) elements of the output stream have been formed—\((n+1)\)-tuples, each of which consists of \(n+1\) elements of the streams \(P_j,\ j=0,1,\ldots,n\), one from each \(P_j\). At the same time queues have formed of \(\bar k_j=k_j-\bar k\) elements of the stream \(P_j\). All \(\bar k_j\ge 0,\ \min_j \bar k_j=0\). We shall say: the state vector \((k_j),\ j=0,1,\ldots,n\), of the system has generated the queue vector \((\bar k_j)\). Likewise, the queue vector gener-
are vector states \((k_j+m)\), \(m\) an arbitrary integer. We form the vector \(K=[\bar k_j]=[k_j]=[k_j+m]\subset \Omega_n\). The components \(\bar k_j\) of the vector-queue \((k_j)\) are the second canonical b.c., while the vector states generating it are the various integer b.c. of this vector. Therefore we shall say: a vector-queue \(K=[\bar k_j]=[k_j]\) has formed; the probabilities \(P(K,t)\) of formation at time \(t\) of such a queue are equal \(\left(\sum_m=\sum_{m=1}^{\infty}\right)\):
\[ P(K,t)=\sum_m\prod_j p_j(k_j+m,t) =\exp\left(-\sum_j \lambda_j t\right)U_K^{(n)}(\lambda_j t), \tag{4} \]
where
\[ U_K^{(n)}(x_j)=U_{(k_j)}^{(n)}(x_j) =U_{k_0,k_1,\ldots,k_n}^{(n)}(x_0,x_1,\ldots,x_n)= \]
\[ =U_{(k_j+m)}^{(n)}(x_j) =\sum_m\prod_j \frac{x_j^{m+k_j}}{(m+k_j)!}. \tag{5} \]
For \(x_0=x_1=\cdots=x_n=x\) we shall have a function of one variable \(x\):
\[ U_K^{(n)}(x)=U_{k_0,k_1,\ldots,k_n}^{(n)}(x) =\sum_m\prod_j \frac{x^{m+k_j}}{(m+k_j)!} =\sum_m \frac{x^{(n+1)m+k}}{\prod_j (m+k_j)!}, \qquad k=\sum_j k_j. \]
For \(n=1\), \(U_{k_0,k_1}^{(1)}(x)=J_{k_1-k_0}(2x)\).
The functions introduced are only loosely connected with Akimov’s generalized Bessel functions \((^2)\).
3. Generating function. Let \(\prod_j t_j=1\); for \(K=[k_j]\), \(t^K=\prod_j t_j^{k_j}\); then
\[ \exp\left(\sum_j t_j x_j\right) = \sum_{K\subset \Omega_n} t^K U_K^{(n)}(x_j). \tag{6} \]
From this follow the addition theorem (7) and formulas (8)—(10):
\[ U_K^{(n)}(x_j+y_j) = \sum_{K'\subset \Omega_n} U_{K'}^{(n)}(x_j)\,U_{K-K'}^{(n)}(y_j); \tag{7} \]
\[ \frac{\partial}{\partial x}U_K^{(n)}(x) = \frac{1}{n+1}U_{K-e_j}^{(n)}(x); \tag{8} \]
\[ \sum_j U_{K-e_j}^{(n)}(x_i)\,x_j e_j = U_K^{(n)}(x_i)\,K; \tag{9} \]
\[ \sum_j x_j e_j = \sum_{K\subset \Omega_n} K U_K^{(n)}(x_i). \tag{10} \]
Let \(Y\subset [y_j]=[y'_j]\), \(K=[k_j]\); \(KY=\sum_j k_j y_j\); \(a_K(Y)=\exp(2\pi iKY)\). The system \(\{a_K(Y)\}_{K\subset\Omega_n}\) is orthogonal in \(Y\) in \(W_n\), complete in the class of functions periodic with periods \(e_j\) and with integrable squares on \(W_n\). Every such function \(f(Y)\) is representable by the Fourier series
\[ f(Y)=\sum_{K\subset\Omega_n} c_K a_K(Y), \qquad c_K=\frac{1}{|W_n|}\int_{W_n}\cdots\int f(Y)a_K(Y)\,d\omega_n. \tag{11} \]
Putting in (6) \(t_j=\exp(2\pi i y'_j)\), \(\sum_j y'_j=0\), \(Y=[y'_j]\), we obtain:
\[ \exp\left(\sum_j x_j \exp(2\pi i y'_j)\right) = \sum_{K\subset\Omega_n} U_K(Y)a_K(Y). \tag{12} \]
Hence, from (11) follows the integral representation
\[ U_{(k_j)}^{(n)}(x_j)=U_K^{(n)}(x_j)= \frac{1}{|W_n|}\int_{W_n}\cdots\int \exp\left[\sum_j\left(\exp(2\pi i y_j')-2\pi i k_j y_j\right)\right]\,d\omega_n . \tag{13} \]
4. Asymptotics. Let \(x\gg 1\), \(k_j=O(\sqrt{x})\), \(K=[k_j]\subset \Omega_n\),
\[ U_K^{(n)}(x)= \frac{1}{\sqrt{(n+1)(2\pi x)^n}} \exp\left[-\frac{n\|K\|^2}{2(n+1)x}\right] \left[1+O\left(\frac{1}{\sqrt{x}}\right)\right]. \tag{14} \]
Let
\[ x\gg 1;\qquad \lambda_j>0;\qquad z_j=\lambda_j-k_jx=O(\sqrt{x}); \]
\[ S_k=\sum_{i=0}^{k}\frac{1}{\lambda_i},\qquad k=0,1,\ldots,n;\qquad z^{(k)}=\frac{1}{S_{k-1}}\sum_{i=0}^{k-1}\frac{z_i}{\lambda_i}, \qquad k=1,2,\ldots,n+1; \]
\[ \Phi(z_j)= \sum_j \frac{z_j^2}{\lambda_j} -\frac{1}{S_n}\left(\sum_j\frac{z_j}{\lambda_j}\right)^2 \equiv \sum_j\frac{1}{\lambda_j}\left(z_j-z^{(n+1)}\right)^2 \equiv \]
\[ \equiv \sum_{k=1}^{n}\frac{S_{k-1}}{\lambda_k S_k} \left(z_k-z^{(k)}\right)^2; \]
\[ U_K^{(n)}(z_j)= \left[\prod_j \lambda_j/S_n(2\pi x)^n\right]^{1/2} \exp\left[-\frac{1}{2x}\Phi(z_j)\right] \left[1+O\left(\frac{1}{\sqrt{x}}\right)\right]; \tag{15} \]
\[
O\left(\frac{1}{\sqrt{x}}\right)\sim
\sum_{p=1}^{\infty}\frac{a_p}{(\sqrt{x})^p};
\]
\(a_p\) is a polynomial of degree \(p\) in \(k_i/\sqrt{x}\) in (14), and in \(z_i/\sqrt{x}\) in (15).
Moscow State University
named after M. V. Lomonosov
Received
15 IX 1967
REFERENCES
¹ Kh. Sh. Margulis, Collection Cybernetics in the Service of Communism, 2, 1964, p. 266. ² M. I. Akimov, On Bessel Functions of Many Variables and Their Applications in Mechanics, Petrograd, 1922.